Generative artificial intelligence is increasingly involved in recommendation, decision making, planning, and task execution, extending human-AI interaction beyond well-defined commands to more open-ended tasks. In these settings, user needs, preferences, and goals are not always fully specified at the outset, but may become clearer or change as users acquire information and receive feedback. Multi-turn interaction and long-term memory further make intent understanding dependent on conversational context and task state. These developments challenge the traditional assumption that user intent is a stable target recoverable from a single utterance. Relevant research spans natural language processing, information retrieval, information behavior, cognitive and decision research, conversation analysis, and human-computer interaction, yet an integrated account remains limited.

This review distinguishes related concepts, including ambiguity, underspecification, clarification, intent elicitation, and intent discovery. Ambiguous intent is defined as a state in which currently available information is insufficient for an AI system to establish adequate task-relevant understanding and respond appropriately. The literature is organized around three interconnected perspectives: language understanding and intent recognition, user cognition and intent formation, and interactive coordination and common ground. On this basis, we propose an expression--construction--coordination (ECC) process-oriented integrative perspective. Expression concerns how existing or partially formed intentions are externalized and understood; construction concerns how user goals and preferences emerge and are revised through information acquisition and interaction; and coordination concerns how users and AI systems establish and repair shared understanding of conversational context and task state. These processes are interrelated rather than mutually exclusive and may overlap across multi-turn interaction.

The reviewed literature suggests that ambiguous intent cannot be understood as a purely linguistic problem. Research on task-oriented dialogue, conversational search, and open-domain question answering has mainly examined how systems identify and recover relatively stable user goals when expressions are incomplete or open to multiple interpretations. Information behavior and decision research, however, shows that users do not always possess fully formed goals before interaction: information seeking, available alternatives, comparison dimensions, and system feedback may contribute to the formation and revision of goals and preferences. Research on conversational repair and common ground further indicates that successful intent understanding depends on whether users and systems maintain compatible representations of shared context and current task state. Taken together, user intent should not always be treated as a static latent variable waiting to be identified, but may instead be expressed, formed, adjusted, and jointly understood through ongoing interaction.

Several directions warrant further investigation from the ECC perspective. First, research should move beyond surface manifestations of ambiguity to examine how user states, task characteristics, and interactional conditions contribute to ambiguous intent. Naturalistic interaction data and controlled experiments may provide complementary evidence. Second, multi-turn research should distinguish increases in system knowledge about an existing goal from genuine changes in the user's goal or preference. Third, interaction support should be studied as context dependent rather than treating clarification as the default response to unclear requests. Future work should compare different forms of support and examine how task risk, action reversibility, interaction cost, and user control influence their appropriateness. Finally, long-term interaction raises questions about how remembered information, users' expectations of system memory, and common ground shape current intent understanding.

These issues also have implications for the design and evaluation of generative AI systems. Ambiguous input should not automatically trigger additional questioning, because appropriate support depends on the task and interaction rather than on linguistic uncertainty alone. Evaluation should extend beyond the relevance or fluency of individual responses to whether an interaction improves task understanding and progress while controlling interaction cost and maintaining user agency, trust, and satisfaction. In long-term interaction, historical information can support continuity and reduce repetitive communication, but only when it remains relevant, is appropriately retrieved, and can be updated when necessary. A process-oriented understanding of ambiguous intent may therefore contribute to both theories of user intent and the design of AI systems that better support open-ended, multi-turn, and long-term interaction.